Résumé
This manuscript presents my research work, which falls within the general field of decision aid. While the spectrum of decision aid is very broad, the work in question focuses on some specific aspects. More precisely, I consider the context where decision problems are related to actions whose consequences may be uncertain, multi-criteria in nature, and involve several decision-makers. Most of the time, the underlying model consists in constructing a utility (or value) function that plays the role of the objective function of a constrained optimization problem to be solved. The targeted applications concern the resolution of ranking or classification problems of alternatives (or classes) that are potential solutions.The ranking problem is addressed as a multiple criteria decision analysis (MCDA) problem, either through methods derived from multi-attribute utility theory (MAUT) or through outranking methods. The second classification problem, although it can be handled by MCDA methods (e.g., Electre Tri), is generally treated as a decision problem under uncertainty using machine learning techniques to approximate posterior probability distributions. In both situations, the decision-maker is asked to provide his preferences, when possible, in order to construct the utility function and compare potential solutions to the problem under consideration.Multiple criteria decision aid, with its two branches—multi-objective optimization (MOO) and multiple criteria decision analysis—aims to assist decision-makers in dealing with complex problems: multiplicity of antagonistic or even contradictory viewpoints, multiplicity of decision-makers with different objectives, algorithmic complexity, difficulty in obtaining information (in terms of cost and time), etc. In particular, we are interested in developing models that allow decision-makers to evaluate the current situation and obtain efficient improvement directions. These models require strong involvement of decision-makers/experts to optimize the efficiency of the solutions. This is why preference learning and soft computing, which facilitate this task for decision-makers, have been of particular interest in our work. Preference learning enables us to determine the set of parameters defining a multicriteria method, while soft computing allows us to represent the behavioral models of complex systems at lower cost.We have contributed to several themes of multiple criteria analysis. Our contributions to multicriteria evaluation consist in proposing or adapting fusion or aggregation operators to particular situations where information comes from multiple sources; information is imprecise and uncertain; criteria/attributes are interactive; decision-makers are multiple; and preference information is difficult to obtain directly. Our contributions to establishing key indices in multicriteria decision-making can be summarized in three points: 1) Decision power indices, which accelerate convergence in collective decision-making; 2) worth index, which help identify the most beneficial criteria to improve as a priority; 3) Contribution indices, which explain decisions by highlighting the dimensions that contribute to their good or poor ranking. The applications concerned by these contributions include preliminary product design, environmental impact assessment (LCA), industrial ecology and risks, and the comparison of low-carbon energy sources.On the other hand, regarding machine learning, we have mainly worked on supervised classification and association rule extraction. More specifically, we have contributed to confidence (or robust) prediction techniques. By confidence, we mean predictions (or rules) that take into account imperfections in the input data to provide cautious outputs. These prediction techniques are particularly useful in sensitive domains such as medical diagnosis, where imperfections are present, as they allow us to represent uncertainty and limit risks accordingly. The general idea is to prioritize the reliability of information rather than the precision of its content. Our work on imprecise classification—i.e., when the output of the classification method for an input sample subject to severe imperfections is a subset of classes—and on imprecise rule extraction—i.e., when an antecedent may have multiple consequences—falls within this framework.For all these contributions, we adopted belief function theory as it allows us to represent both imprecision and uncertainty and offers a very interesting framework for combining information. We also proposed, within the belief function framework, extensions of certain operations to the case of proportion vectors. We applied this work to the context of environmental compliance, such as the problem of sorting plastics for recycling, where identification errors can cause major recycling difficulties and significantly degrade the performance of secondary raw materials, but also in the management of humanitarian missions, where mission effectiveness must not be disrupted by lack of information.